WorksComputer Science
Meta-RTL: Evolving Agent Harnesses for RTL Optimization
Abstract
Large language models enable agents to optimize RTL using EDA feedback, yet their executable search procedures remain largely hand- engineered. We present Meta-RTL, an EDA-grounded framework that evolves RTL optimization harnesses from a minimal loop while keeping the optimization model, EDA flow, and per-design budgets fixed. Compared with Dr. RTL, Meta-RTL* reduces aggregate WNS/TNS violations by 7.8%/11.0% across 20 development designs, with area reductions on 70% of them. Without further evolution, Meta-RTL* transfers to 10 unseen IC-RTL designs, reducing aggregate WNS/TNS violations by 20.5%/14.0% and total synthesized area by 12.6%. We further characterize three search mechanisms in the evolved harness and assess their contributions through ablations and a case study.
Introduction
Optimizing register-transfer-level (RTL) designs for timing and area is a critical task in digital hardware design. Functionally equivalent RTL descriptions can synthesize into substantially different circuits, so the way a design is written directly determines its timing and area. Synthesis tools [1], [2] apply predefined optimization passes and heuristics, but they optimize within the structure that the RTL expresses rather than changing it. Designers therefore restructure RTL manually, with each revision followed by functional verification, synthesis, and timing–area analysis. This iterative process is labor-intensive and difficult to scale as designs grow more complex.
Large language models (LLMs) have opened new opportunities to automate this process. Early studies generated RTL from high-level specifications, improving generation quality through domain-specific fine-tuning and functional validation [3, 4, 5]. Later approaches fed compilation, simulation, and PPA results back to the model to iteratively refine generated designs [6], [7]. Recent agentic frameworks go further, coupling LLM reasoning with EDA tools to generate, verify, and optimize RTL designs autonomously [8, 9, 10, 11, 12, 13].
For such agents, how the search is organized matters as much as which transformations the LLM proposes. Every candidate must remain functionally equivalent, timing and area often conflict, and each evaluation requires synthesis and equivalence checking. The search is governed by the agent’s harness: the executable procedure that presents EDA feedback to the model, decides which candidates to keep, and allocates the evaluation budget. It therefore shapes both optimization quality and cost [14]. Existing agents differ precisely in these choices, from predefined workflows for tool invocation and iterative rewriting [9, 10, 11] to persistent memory, reusable skills, and shared candidate pools [12], [13], [15]. Yet all of these harnesses are designed by hand, and each embeds fixed assumptions about how optimization should proceed. Promoting only the best-scoring candidate, for example, assumes that good designs are reached through steady improvement, whereas profitable restructuring often passes through intermediate designs that look worse before they are refined. In these agents, EDA feedback improves the designs but not the harness that produces them.
This raises a natural question: Without manually prescribing sophisticated search heuristics, can we evolve RTL optimization harnesses from EDA feedback that outperform hand-designed ones—and if so, what makes the evolved harnesses effective?
To answer these questions, we present Meta-RTL, which evolves an RTL optimization harness from a minimal seed through two nested loops. In the inner loop, a fixed LLM optimizes RTL designs under the current harness. In the outer loop, a proposer LLM improves the harness much as an engineer would: it traces EDA failures in the inner-loop logs back to specific harness decisions, hypothesizes a targeted fix and its expected effect, probes the fix on a few relevant designs, and revises it based on the evidence. Each submitted harness is then frozen and scored on all development designs by synthesis and equivalence checking outside its control. In this way, EDA feedback improves not only the designs but also the harness that produces them. While skill libraries capture what transformations to apply, Meta-RTL learns how to search for effective transformations, turning a traditionally hand-designed process into an optimization target.
The main contributions of this work are summarized as follows:
We formulate EDA-grounded harness evolution for RTL timing–area optimization, in which the search procedure itself evolves from a minimal loop under a fixed LLM, EDA flow, and budget per design. Its proposer harnesses through four designed stages described in § 3-C.
We show that harness evolution yields substantial timing and area gains (Figure 1). Compared with Dr. RTL, Meta-RTL* reduces aggregate WNS/TNS violation magnitudes by 7.8%/11.0% and area on 14 of 20 development designs (Section IV-B). Without further evolution, it also improves WNS, TNS, and area on 7, 8, and 7 of the 10 unseen designs, respectively (Section IV-C).
We identify three search mechanisms that emerge in the evolved harness, and further assess their contributions through targeted ablations and a circuit-level case study (Sections IV-E and IV-F).

Figure 1. Meta-RTL* versus Dr. RTL on 20 development and 10 unseen designs. Triplets show relative reductions in WNS/TNS violations and area; blue/orange indicates improvement/regression.
Related Work
Conventional Synthesis and RTL Optimization
Traditional hardware optimization relies on logic synthesis and physical design tools to improve area and timing. Early synthesis systems such as SIS [16] introduced systematic logic and sequential circuit optimization. Yosys [1] provides an extensible framework for RTL synthesis, while ABC [2] supports AIG-based logic optimization and technology mapping. OpenROAD [17] enables automated physical design through placement, routing, and timing optimization. These tools employ well-established optimization algorithms, but high-level RTL restructuring still relies on manual design expertise.
LLM-Based RTL Optimization
Recent studies leverage LLMs for iterative RTL optimization, in which candidate designs are repeatedly rewritten and evaluated with EDA feedback [10], [11], [18]. Dr. RTL [12] introduces reusable optimization skills distilled from tool feedback. RTLScout [13] instead exploits shared candidate pools and experience accumulated across runs. Despite these advances, the executable procedures organizing optimization remain largely hand-engineered.
Building on broader progress in self-evolving systems [19, 20, 21], recent studies have explored program evolution for EDA tasks. These include logic synthesis [22], technology mapping [23], and physical design tasks such as placement [24], [25], routing [26], and FPGA place-and-route [27]. Similar approaches have also been applied to hardware design [28] and verification [29].
Unlike approaches that evolve internal EDA heuristics, Meta-RTL evolves the executable search procedure of RTL optimization agents. We investigate whether synthesis and equivalence-checking feedback can guide this evolution to outperform hand-engineered agents and generalize to unseen designs.
Meta-RTL
Overview
As shown in Figure 2, Meta-RTL evolves executable RTL optimization harnesses from a minimal loop that simply edits RTL and evaluates the resulting designs using default EDA reports. An agentic proposer progressively revises the harness using the implementations and execution evidence of earlier rounds, and the evaluation protocol specifies how it interacts with the fixed evaluator.
Figure 2. Overview of Meta-RTL: inner-loop RTL optimization and outer-loop harness evolution.
For an RTL design and resource budget , a harness defines an inner optimization procedure that coordinates a fixed LLM with equivalence checking, synthesis, and timing analysis. Its execution produces a trajectory containing RTL candidates, tool outcomes, timing–area measurements, and search decisions. The harness determines how candidates are constructed, how feedback is presented, and how subsequent transformations are explored.
An archive stores harness implementations, evaluation results, and execution trajectories. Harnesses are ranked by normalized timing–area utility (§ 3-D), and the highest-scoring eligible harness, Meta-RTL, is selected and then evaluated on unseen designs.
Harness Representation and Search Space
A harness is a self-contained Python package that orchestrates the inner optimization loop via two primitives: invoking the fixed LLM and querying the fixed EDA evaluator for synthesis, timing, and sequential equivalence checking (SEC). can freely define prompt templates, helper scripts, and search routines, through which evolution may modify four aspects of the search procedure: (1) candidate construction, (2) feedback organization, (3) search-state management, and (4) resource allocation and search control.
Harnesses execute in Bubblewrap-isolated environments. The independent evaluator, which cannot be accessed or modified by the proposer or , fixes the optimization LLM, EDA flow, synthesis constraints, equivalence checking, scoring objective, and per-design budgets. This isolation ensures a consistent and reliable evaluation protocol while allowing the executable search procedure to evolve. Consequently, neither nor can relax verification constraints or alter scoring logic, so all gains can be attributed to improved search.
Evidence-Guided Harness Evolution
The proposer is an LLM agent that can read the archive , edit harness code, and launch probe runs. At each round, it selects as parent the harness it judges most promising, which need not be the highest-scoring one, based on the implementations, scores, and run records of all archived harnesses. This lets evolution continue from a harness whose new mechanism looks promising in its traces before it pays off in score. Its instructions then prescribe four stages, shown at the bottom of Figure 2, for developing revisions from and its run records .
EDA Failure Attribution. The proposer inspects , including RTL edits, SEC diagnostics, WNS/TNS/area trajectories, and resource usage, to trace how candidate transformations fare in correctness checking and EDA evaluation. It identifies recurring failures and unproductive search behavior, and attributes them to the responsible harness decisions, such as how candidates are constructed or how search state is managed. A score alone cannot provide this information: a harness whose edits fail SEC and one whose valid edits bring no gain may both score zero. The resulting diagnosis names a bottleneck together with the traces that support it.
Harness Hypothesis Generation. The proposer turns the diagnosis into a hypothesis: a targeted change to relevant code paths, prompts, or tool interfaces, paired with a testable prediction of how search behavior should change. Typical predictions include more candidates passing SEC, broader exploration, or fewer unproductive evaluations. Stating the prediction in advance makes a probe informative even when the score barely moves, and keeps a lucky gain from being credited to the change. The proposer then implements it, and the prediction decides which designs to probe and how to read the results.
Cost-Aware Probing. The proposer runs the revised harness on development designs selected for their relevance to the bottleneck, their ability to test the prediction, and their evaluation cost. Because has already run on the same designs, comparing the new traces with isolates the effect of the change and shows whether the predicted behavior occurred, for example in candidate validity, optimization progress, or resource usage. Each candidate is limited to probe runs. Since a full-set evaluation runs every development design under budget , these probes give an early signal at a fraction of its cost.
Evidence-Based Revision. The proposer uses the probe results to accept, refine, or replace the hypothesis. Confirmed improvements are developed further, while failed predictions and regressions lead to a revised implementation or a new hypothesis. Failed predictions are also informative, since they rule out a suspected cause and sharpen the next diagnosis. The proposer may cycle through hypothesis generation, probing, and revision within the -run limit, but each candidate must end in a submitted harness.
Each round produces up to candidate harnesses. Every submitted harness is frozen and formally evaluated on the full development set under budget to obtain , and only harnesses that complete this evaluation are eligible for selection. The archive records every harness together with its lineage, score, and run records, so later rounds build on all accumulated evidence. After rounds, the eligible harness with the highest is selected as Meta-RTL*.
EDA-Grounded Harness Evaluation
Harness evaluation measures the RTL quality that a complete optimization run achieves under fixed resource and EDA constraints. Each run starts from the original design . The returned RTL is eligible for scoring only if it passes correctness checking and yields valid synthesis measurements. The original design remains available as a fallback with zero normalized utility.
Normalized Timing–Area Utility. Meta-RTL uses an equivalent formulation of the timing–area score in Dr. RTL [12]. For a completed run , the normalized gain in each metric is defined as
where WNS, TNS, and denote worst negative slack, total negative slack, and area, respectively. Run utility is defined as
, , and weight the normalized gains, while penalizes relative area growth exceeding . Larger utility indicates greater timing–area optimization gains relative to the original design.
Harness quality is the mean development-set utility:
Normalization measures improvement relative to each original design, while averaging summarizes performance across the development set. All methods use the same scoring parameters (Section IV-A).
Resource-Constrained Execution. Each inner optimization run receives a budget , enforcing
where counts candidate EDA evaluations and denotes cumulative LLM API cost, accounting for cached inputs, uncached inputs, and output tokens. The shared evaluator enforces both limits independently of harness decisions.
Experiments
Experimental Setup
We compare Meta-RTL* with two strong LLM-based RTL optimization frameworks: Dr. RTL [12] with its released skill library and RTLScout [13]. For RTLScout, we only evaluate its RTL optimization component while keeping the synthesis flow fixed.
Evaluation Benchmarks. Our experiments use two benchmark suites, Dr. RTL [12] and IC-RTL [11]. The Dr. RTL benchmark contains 20 designs covering arithmetic, control, communication, and sequential logic, which serve as the development set. IC-RTL contains 10 more complex designs spanning image processing, compression, geometric computation, and hardware acceleration. These designs are held out from harness evolution to evaluate generalization.
Outer Evolution. Evolution begins with a minimal seed: an in-place RTL editing loop guided by default EDA reports. GPT-5.6 Sol [30] serves as the harness proposer, generating up to candidates per round. We conduct evolution rounds, yielding 16 submitted harnesses in total. Each candidate is limited to probe runs on proposer-selected designs. Submitted harnesses are evaluated afresh on all 20 development designs under the fixed inner budget. The highest-scoring eligible harness is selected as Meta-RTL*.
Inner Optimization. All methods share the same EDA flow and use DeepSeek-V4.1-Flash [31] as the optimization model. Synopsys
| Design | Dr. RTL WNS (ns) | Dr. RTL TNS (ns) | Dr. RTL () | Dr. RTL | RTLScout WNS (ns) | RTLScout TNS (ns) | RTLScout () | RTLScout | Meta-RTL* WNS (ns) | Meta-RTL* TNS (ns) | Meta-RTL* () | Meta-RTL* |
| DSP | -2.59 | -141.81 | 4127.8 | 0.0527 | -2.57 | -139.78 | 4162.4 | 0.0600 | -2.56 | -126.17 | 3996.7 | 0.0979 |
| FIFO | -0.40 | -1628.64 | 35613.1 | 0.1968 | -0.41 | -1626.13 | 34519.1 | 0.1925 | -0.26 | -942.06 | 37448.5 | 0.4388 |
| LSTM | -3.09 | -81.84 | 3641.3 | 0.5541 | -3.19 | -89.39 | 3863.9 | 0.5283 | -2.59 | -70.72 | 4972.3 | 0.6023 |
| SPI | -0.24 | -24.88 | 1916.0 | 0.2184 | -0.32 | -25.49 | 1490.4 | 0.1224 | -0.22 | -22.20 | 1849.5 | 0.2819 |
| UART | -0.35 | -22.13 | 1134.0 | 0.0714 | -0.37 | -22.48 | 1147.0 | 0.0383 | -0.35 | -21.36 | 1001.2 | 0.0987 |
| aes | -0.61 | -789.44 | 34532.7 | 0.1084 | -0.70 | -913.49 | 33755.1 | 0.0000 | -0.52 | -684.54 | 31071.2 | 0.2282 |
| arm_cpu1 | -5.18 | -2174.25 | 22520.4 | 0.1117 | -5.18 | -2846.87 | 22577.8 | 0.0365 | -5.17 | -2224.34 | 22392.1 | 0.1079 |
| arm_cpu2 | -0.84 | -562.75 | 11495.2 | 0.1256 | -0.91 | -603.82 | 11098.3 | 0.0748 | -0.86 | -572.92 | 11715.7 | 0.1072 |
| communication | -0.36 | -70.71 | 1901.4 | 0.0750 | -0.38 | -71.95 | 1870.2 | 0.0463 | -0.20 | -44.50 | 2005.1 | 0.3931 |
| controller | -0.38 | -2.85 | 234.6 | 0.0000 | -0.31 | -2.46 | 231.7 | 0.1419 | -0.17 | -2.34 | 232.2 | 0.3405 |
| cpu_fsm | -0.57 | -425.66 | 32069.0 | 0.1567 | -0.61 | -415.03 | 32256.5 | 0.1401 | -0.50 | -433.57 | 32053.3 | 0.1930 |
| cpu_pipe | -2.31 | -260.11 | 5478.0 | 0.0393 | -2.37 | -277.42 | 5349.5 | 0.0094 | -2.31 | -257.07 | 5206.4 | 0.0505 |
| datapath | -0.83 | -474.59 | 12361.6 | 0.0521 | -0.88 | -513.42 | 12136.8 | 0.0000 | -0.81 | -472.61 | 12281.8 | 0.0658 |
| i2c | -0.31 | -24.98 | 1297.0 | 0.0908 | -0.34 | -25.42 | 1254.5 | 0.0483 | -0.31 | -24.52 | 1255.8 | 0.1016 |
| pcie | -0.57 | -20.25 | 1037.4 | 0.2697 | -0.52 | -20.30 | 1006.8 | 0.3027 | -0.35 | -17.06 | 1036.6 | 0.4558 |
| router | -0.38 | -218.76 | 5530.7 | 0.2260 | -0.48 | -271.01 | 5598.8 | 0.0667 | -0.38 | -221.17 | 5432.0 | 0.2258 |
| simple_spi | -0.30 | -31.88 | 1227.1 | 0.0450 | -0.32 | -32.02 | 1194.1 | 0.0164 | -0.27 | -31.06 | 1210.3 | 0.1025 |
| ticket_machine | -0.10 | -0.29 | 31.9 | 0.6395 | -0.14 | -0.38 | 29.5 | 0.5318 | -0.08 | -0.45 | 47.1 | 0.6088 |
| tv80 | -1.22 | -362.94 | 5999.9 | 0.0522 | -1.25 | -366.76 | 5980.5 | 0.0377 | -1.10 | -342.05 | 6595.5 | 0.1024 |
| vending_machine | -0.08 | -0.16 | 20470.6 | 0.6471 | -0.09 | -0.16 | 20514.5 | 0.6282 | -0.09 | -0.24 | 12896.2 | 0.6566 |
| Mean | – | – | – | 0.1866 | – | – | – | 0.1511 | – | – | – | 0.2630 |
Table 1. Timing and area comparison on the 20 Dr. RTL development designs. Bold indicates the best unrounded result.
Design Compiler [32] is used for synthesis and timing analysis, and Cadence JasperGold [33] for SEC. To assess RTL-level improvements beyond standard synthesis optimizations, we synthesize with a commercial tool and the Nangate 45 nm [34] standard-cell library. Following Dr. RTL [12], we impose a stringent 0.1 ns clock period to drive aggressive timing optimization. All methods share per-design budgets of EDA evaluations and in LLM API cost. The 50-evaluation cap matches Dr. RTL’s 10 major rounds of five minor attempts each.
Evaluation Metrics. We report worst negative slack (WNS), total negative slack (TNS), synthesized area (), and normalized timing–area utility (). The utility metric, defined in § 3-D, uses , , and , following Dr. RTL [12]. Higher utility indicates better timing–area optimization.
Computational Cost. Under DeepSeek-V4.1-Flash pricing, the estimated costs of optimizing all 20 development designs are US$22.5 for Meta-RTL*, US$75 for Dr. RTL, and US$12 for RTLScout. The total recorded LLM API expenditure was US$627.09, excluding infrastructure costs (e.g., EDA licenses).
The candidate harnesses yield an estimated evolution cost of US$510 (16US$22.5, i.e., Meta-RTL*’s cost per candidate, plus US$150 for GPT-5.6 Sol proposals). Amortizing this cost over later applications, Meta-RTL* would become more cost-efficient than Dr. RTL after about 200 design optimizations, even ignoring the effort behind the hand-engineered baselines.
Development-Set Performance
Table I compares Meta-RTL* with Dr. RTL and RTLScout on the 20 development designs. Among the three methods, Meta-RTL* achieves the best or tied-best WNS on 18 designs, the best TNS on 14 designs, and the lowest area on 8 designs.
The results show a strong advantage in timing. Compared with RTLScout, Meta-RTL* achieves better WNS on 19 designs, ties on one, and improves TNS on 17 designs, while RTLScout obtains lower area on 12 designs. Timing gains also coincide with area reductions: compared with Dr. RTL, Meta-RTL* improves or maintains both WNS and TNS while reducing area on 10 designs, including 7 designs with strict improvements in all three metrics.
| Design | Dr. RTL WNS | Dr. RTL TNS | Dr. RTL | Dr. RTL | Meta-RTL* WNS | Meta-RTL* TNS | Meta-RTL* | Meta-RTL* |
| LBP | -0.26 | -22.30 | 1171 | 0.2810 | -0.20 | -20.41 | 1092 | 0.3866 |
| CONV | -2.95 | -165.59 | 4103 | 0.0909 | -2.86 | -150.58 | 4166 | 0.1307 |
| HC | -0.62 | -111.54 | 2996 | 0.0552 | -0.63 | -112.25 | 2962 | 0.0475 |
| JAM | -0.79 | -46.67 | 1599 | 0.1702 | -0.40 | -34.46 | 1737 | 0.4462 |
| DT | -0.65 | -35.57 | 1594 | 0.0439 | -0.39 | -27.32 | 1354 | 0.3326 |
| REF | -6.67 | -1789.23 | 35970 | 0.1726 | -3.62 | -1311.44 | 26768 | 0.4842 |
| GEO | -2.85 | -699.03 | 14976 | 0.0534 | -2.97 | -727.84 | 13207 | 0.0387 |
| IOT | -0.80 | -152.88 | 5448 | 0.3401 | -0.73 | -110.02 | 5070 | 0.4410 |
| SECC | -0.99 | -25.05 | 1835 | 0.1158 | -0.70 | -22.59 | 1920 | 0.2704 |
| GEMM | -2.83 | -1573.07 | 41400 | 0.0651 | -2.93 | -1455.20 | 38186 | 0.0819 |
| Mean | – | – | – | 0.1388 | – | – | – | 0.2660 |
Table 2. Timing and area comparison on 10 unseen IC-RTL designs
Overall, Meta-RTL* achieves the highest mean normalized utility of 0.2630, compared with 0.1866 for Dr. RTL and 0.1511 for RTLScout, indicating the strongest aggregate timing–area performance.
Generalization to Unseen Designs
To evaluate generalization, we compare Meta-RTL* with Dr. RTL on 10 unseen IC-RTL designs. Both methods leverage optimization experience from the 20-design Dr. RTL benchmark: Dr. RTL through its distilled skill library, and Meta-RTL through harness evolution. For a fair comparison, we exclude RTLScout from the transfer evaluation, as it was not developed on the same benchmark. The skill library and evolved harness are kept fixed during evaluation on IC-RTL.
Table II presents the timing and area results. Meta-RTL* achieves better WNS on 7 designs, better TNS on 8, and lower area on 7. It also achieves higher utility on 8 of the 10 designs. Notably, on REF, Meta-RTL* simultaneously improves WNS, TNS, and area by 45.7%, 26.7%, and 25.6%, respectively. Overall, its mean utility reaches 0.2660, compared with 0.1388 for Dr. RTL, representing a 91.6% relative improvement. Note that REF contributes substantially to the aggregate timing gains, yet Meta-RTL* still achieves 79.0% higher mean utility on the remaining nine designs. These results indicate that the evolved harness transfers effectively to unseen designs.
Evolution Analysis
Evolutionary Progress. Figure 3 traces the evolution of Meta-RTL{sup}*. Starting from the minimal loop, Local Feedback extracts critical-path information and achieves a utility of 0.0936. Serial Dual Context then improves utility to 0.1265 by revisiting timing-improved candidates for area recovery.
Figure 3. Evolution lineage of Meta-RTL harnesses.
The subsequent evolution exemplifies how the proposer translates execution evidence into harness revisions. Inspecting Serial Dual Context trajectories, the proposer identified repeated RTL editing failures and premature search termination that prevented structural transformations from reaching EDA evaluation. It diagnosed a delivery/continuation bottleneck and hypothesized that programmatic RTL transformations and bounded search continuation would enable more valid measurements. Targeted probing on four development designs increased mean utility from 0.0817 to 0.2263, with previously unsuccessful transformations passing SEC. The resulting Programmatic Relay achieved a full-set utility of 0.2332.
Further evolution revealed the overhead of repeatedly regenerating transformation programs, motivating persistent artifacts and incremental program edits. Although Adaptive Workbench temporarily dipped to 0.2060 due to unpruned workspace state across retries, selecting it as the parent allowed the proposer to isolate this bottleneck and introduce structured artifact indexing, yielding Artifact Workbench (, selected as Meta-RTL{sup}*).
Discovered Mechanisms. The lineage reveals three complementary search mechanisms in Meta-RTL{sup}*.
EDA-Grounded Transformation Guidance. The evolved harness links EDA feedback to targeted structural transformations. By interpreting critical-path evidence and timing–area measurements, it guides the LLM toward transformations such as arithmetic sharing, control simplification, and common-logic factoring. Subsequent EDA results then refine the diagnosis, closing an evidence-guided loop.
Executable Transformation Workspace. The evolved harness maintains executable transformation programs with their intermediate artifacts and execution feedback across model requests. Rather than regenerating RTL edits from scratch, the optimizer can inspect, revise,
| Config. | LBP | CONV | HC | JAM | DT | REF | GEO | IOT | SECC | GEMM | Mean |
| Full | 0.387 | 0.131 | 0.048 | 0.446 | 0.333 | 0.484 | 0.039 | 0.441 | 0.270 | 0.082 | 0.266 |
| −Guidance | 0.299 | 0.141 | 0.043 | 0.436 | 0.243 | 0.238 | 0.056 | 0.407 | 0.153 | 0.047 | 0.206 |
| −Frontier | 0.322 | 0.095 | 0.048 | 0.413 | 0.263 | 0.130 | 0.014 | 0.418 | 0.150 | 0.035 | 0.189 |
| −Workspace | 0.362 | 0.027 | 0.038 | 0.308 | 0.316 | 0.122 | 0.014 | 0.385 | 0.124 | 0.012 | 0.171 |
Table 3. Ablations on 10 unseen designs (bold: best unrounded)
and rerun them. This enables incremental construction and debugging of complex structural rewrites, including repairs of failed attempts.
Pareto Frontier Memory. The evolved harness retains verified RTL candidates representing different trade-offs among WNS, TNS, and area. Rather than committing to a single incumbent, the optimizer can revisit promising intermediate candidates and explore alternative optimization directions. This preserves opportunities for further timing or area improvements that might otherwise be lost during the search.
Ablation Studies
Effect of Harness Evolution. To assess the benefit of harness evolution, we compare Meta-RTL with direct harness design. Using the same proposer, 16 harnesses are independently developed from the same minimal seed, each with up to probe runs but no access to other candidates or their execution histories. All are evaluated under the same protocol, and the best is further tested on unseen designs.
Figure 4 shows the utility of the 16 independently designed harnesses. The best achieves 0.1813, below both Dr. RTL (0.1866) and Meta-RTL{sup}* (0.2630). Their mean utility is 0.1376. On unseen designs, the selected harness achieves 0.1050, compared with 0.2660 for Meta-RTL{sup}*. These results highlight the benefit of accumulating and organizing execution evidence across harness revisions.
Figure 4. Development-set utility of 16 independently designed harnesses.
Contributions of Evolved Mechanisms. To assess the contributions of the three discovered mechanisms, we construct three ablated variants of Meta-RTL{sup}* by individually disabling EDA-grounded transformation guidance (−Guidance), Pareto frontier memory (−Frontier), and the executable transformation workspace (−Workspace). All variants are evaluated on the same 10 unseen IC-RTL designs under identical LLM, EDA flow, and budgets, with other components unchanged.
Table III reports the normalized timing–area utility. Removing Guidance, Frontier, and Workspace reduces the mean utility from 0.2660 to 0.2064, 0.1889, and 0.1707, respectively, corresponding to decreases of 22.4%, 29.0%, and 35.8%. The full harness outperforms these variants on 8, 9, and 10 of the 10 designs, respectively.
Workspace removal causes the largest degradation, with lower utility on every unseen design. On REF, removing Workspace or Frontier reduces utility from 0.484 to 0.122 or 0.130, respectively. Removing Workspace also lowers utility on CONV from 0.131 to 0.027. These results demonstrate the complementary roles of retaining executable transformations for iterative refinement and preserving promising RTL candidates for continued exploration.
A Case Study
Figure 5 illustrates, on the controller design, how the discovered mechanisms interact within the final Meta-RTL{sup}*` harness.
Figure 5. Case study of controller optimization. TNS is omitted for clarity, as it did not affect the key trade-offs illustrated here.
Dr. RTL employs repeated five-way candidate expansion and single-candidate promotion, but its search remains focused on expanded decoding without exploring alternative structural directions. Despite evaluating 50 subsequent candidates, it fails to overcome the area penalty, leaving the original v0 as the best design. RTLScout carries the best design from each run into subsequent runs, but abandons the promising one-hot encoding in v14, ultimately settling for a suboptimal four-bit encoding without further refinement.
Through harness evolution, Meta-RTL{sup}*` gains the ability to identify, preserve, and revisit promising intermediate designs. The evolved harness maintains a Pareto frontier of designs, restores their RTL sources across optimization contexts, and exposes executable interfaces for reusing saved transformation programs. These mechanisms make intermediate designs reusable starting points for structural refinement, even when their immediate optimization scores worsen.
Here, v6 uses a two-bit group/two-bit index encoding to reduce worst-path logic depth from eight to five levels, but retains per-state output decoding with a 14.2% area overhead. The harness preserves v6 in frontier memory, excludes the dominated v7–v10, and later restores v6 to apply the saved program w51. By exploiting the common two-bit prefix of four states asserting bypass_rk, w51 replaces their decoding branches with one predicate, reducing the fan-in cone from 23 to 2 cells and total combinational cells from 247 to 188. The resulting v11 passes equivalence checking and reduces area and WNS/TNS violations by 8.2%, 18.4%, and 16.1%, respectively, relative to v0, Dr. RTL’s best result. It also matches RTLScout’s best WNS with 7.0% lower area. As a structural stepping stone, v11 provides factored output logic that subsequent candidates reuse and refine through further state recoding, ultimately yielding v22 (Table I).
Limitations and Future Directions
Our results establish the feasibility and effectiveness of EDA-grounded harness evolution under a fixed proposer model, optimization model, EDA flow, and technology library. However, given the computational budget, the robustness of these findings to LLM stochasticity warrants further evaluation through repeated trials.
Future work could investigate how different proposer and optimization models shape the search strategies discovered through evolution. Evaluating across technology libraries and physical design flows would further examine the transferability of these strategies under different hardware constraints. Beyond RTL optimization, extending harness evolution to other EDA tasks may reveal new mechanisms arising from distinct feedback signals and search spaces.
Conclusion
We presented Meta-RTL, which evolves executable RTL optimization harnesses under a fixed LLM and EDA flow. Compared with Dr. RTL, Meta-RTL{sup}*` reduces aggregate WNS and TNS violations by 7.8% and 11.0% on 20 development designs and transfers without further evolution to 10 unseen designs, where it nearly doubles mean utility (0.2660 vs. 0.1388). Analysis and ablations identify EDA-grounded transformation guidance, an executable transformation workspace, and a Pareto frontier memory as key mechanisms. These findings show that EDA feedback can improve not only RTL designs, but also the search procedures used to optimize them.
Content Generated by AI
ChatGPT (OpenAI) was used to assist with drafting and refining the paper, as well as generating and revising portions of the released codebase. All AI-assisted content was reviewed by the authors, who take full responsibility for the final work.
References
- [1]C. Wolf and J. Glaser, “Yosys – a free Verilog synthesis suite,” in Proceedings of the Australasian Workshop on Microelectronic Systems, 2013, pp. 47–52.
- [2]R. K. Brayton and A. Mishchenko, “ABC: An academic industrial-strength verification tool,” in Computer Aided Verification (CAV), ser. Lecture Notes in Computer Science, vol. 6174. Springer, 2010, pp. 24–40.DOI
- [3]S. Thakur, B. Ahmad, H. Pearce, B. Tan, B. Dolan-Gavitt, R. Karri, and S. Garg, “VeriGen: A large language model for Verilog code generation,” ACM Transactions on Design Automation of Electronic Systems, vol. 29, no. 3, 2024.arxiv.org/abs/2308.00708
- [4]S. Liu, W. Fang, Y. Lu, J. Wang, Q. Zhang, H. Zhang, and Z. Xie, “RTLCoder: Fully open-source and efficient LLM-assisted RTL code generation technique,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 44, no. 4, pp. 1448–1461, 2025.DOI
- [5]A. Wei, H. Tan, T. Suresh, D. Mendoza, T. S. F. X. Teixeira, K. Wang, C. Trippel, and A. Aiken, “VeriCoder: Enhancing LLM-based RTL code generation through functional correctness validation,” arXiv preprint arXiv:2504.15659, 2025.arxiv.org/abs/2504.15659
- [6]S. Thakur, J. Blocklove, H. Pearce, B. Tan, S. Garg, and R. Karri, “AutoChip: Automating HDL generation using LLM feedback,” arXiv preprint arXiv:2311.04887, 2023.arxiv.org/abs/2311.04887
- [7]K. Thorat, J. Zhao, Y. Liu, A. Hasan, H. Peng, X. Xie, B. Lei, and C. Ding, “LLM-VeriPPA: Power, performance, and area optimization aware Verilog code generation with large language models,” in Proceedings of the ACM/IEEE Symposium on Machine Learning for CAD (MLCAD), 2025.arxiv.org/abs/2510.15899
- [8]C.-T. Ho, H. Ren, and B. Khailany, “VerilogCoder: Autonomous Verilog coding agents with graph-based planning and abstract syntax tree (AST)-based waveform tracing tool,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 1, pp. 300–307, 2025.arxiv.org/abs/2408.08927
- [9]K. Tasnia, A. Garcia, T. Farheen, and S. Rahman, “VeriOpt: PPA-aware high-quality Verilog generation via multi-role LLMs,” in Proceedings of the IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2025, pp. 1–9.arxiv.org/abs/2507.14776
- [10]Y. Wang, W. Ye, P. Guo, Y. He, Z. Wang, B. Tian, S. He, G. Sun, Z. Shen, S. Chen, A. Srivastava, Q. Zhang, G. Qu, and A. Li, “Sym-RTLO: Enhancing RTL code optimization with LLMs and neuron-inspired symbolic reasoning,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 38, 2025, pp. 55 809–55 834.arxiv.org/abs/2504.10369
- [11]W.-P. Hsin, R.-H. Deng, Y.-T. Hsieh, E.-M. Huang, and S.-H. Hung, “EvoIVE: Evolutionary search for LLM-based Verilog generation and optimization,” arXiv preprint arXiv:2601.18067, 2026.arxiv.org/abs/2601.18067
- [12]W. Fang, Y. Lu, S. Liu, J. Wang, Z. Guo, J. He, F. Tu, and Z. Xie, “Dr. RTL: Autonomous agent RTL optimization through tool-grounded self-improvement,” in Proceedings of the IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2026.arxiv.org/abs/2604.14989
- [13]F. Arnold, R. Amaudruz, D. Tsaras, R. Andri, and L. Cavigelli, “RTLScout: Joint agentic code and synthesis optimization for efficient digital circuits,” arXiv preprint arXiv:2606.06530, 2026.arxiv.org/abs/2606.06530
- [14]S. Cuyckens, M. Jivanescu, J. Yin, C. Fang, and M. Verhelst, “ARES: Adaptive reasoning-effort steering for PPA- and cost-aware RTL optimization with LLM agents,” arXiv preprint arXiv:2607.27879, 2026.arxiv.org/abs/2607.27879
- [15]Y. Wang, Q. Shi, S. Li, Q. Hu, X. Yin, B. Guo, X. Han, M. Sun, and J. Su, “VeriAgent: A tool-integrated multi-agent system with evolving memory for PPA-aware RTL code generation,” arXiv preprint arXiv:2603.17613, 2026.arxiv.org/abs/2603.17613
- [16]E. M. Semtovick, K. J. Singh, L. Lavagno, C. Moon, R. Murgai, A. Saldanha, H. Savoj, P. R. Stephan, R. K. Brayton, and A. L. Sangiovanni-Vincentelli, “SIS: A system for sequential circuit synthesis,” University of California, Berkeley, Tech. Rep. UCB/ERL M92/41, 1992.
- [17]T. Ajayi, V. A. Chhabria, M. Fogaça, S. Hashemi, A. Hosny, A. B. Kahng, M. Kim, J. Lee, U. Mallappa, M. Neseem, G. Pradipta, S. Reda, M. Saligane, S. Sapatnekar, C. Sechen, M. Shalan, W. Swartz, L. Wang, Z. Wang, M. Woo, and B. Xu, “Toward an open-source digital flow: First learnings from the OpenROAD project,” in Proceedings of the 56th ACM/IEEE Design Automation Conference (DAC), 2019.
- [18]H. Ping, P. Zhang, Z. Wang, S. Li, A. Cheng, W. Yang, P. Bogdan, and S. Nazarizadeh, “POET: Power-oriented evolutionary tuning for LLM-based RTL PPA optimization,” arXiv preprint arXiv:2603.19333, 2026.arxiv.org/abs/2603.19333
- [19]J. Lin, S. Liu, C. Pan, L. Lin, S. Dou, Z. Xi, X. Huang, H. Yan, Z. Han, T. Gui, and Y.-G. Jiang, “Agentic harness engineering: Observability-driven automatic evolution of coding-agent harnesses,” arXiv preprint arXiv:2604.25850, 2026.arxiv.org/abs/2604.25850
- [20]Y. Lee, R. Nair, Q. Zhang, K. Lee, O. Khattab, and C. Finn, “Meta-Harness: End-to-end optimization of model harnesses,” arXiv preprint arXiv:2603.28052, 2026.arxiv.org/abs/2603.28052
- [21]H. Zhang, S. Zhang, K. Li, C. Zhang, Y. Chen, Y. Zhang, L. Bai, and S. Hu, “Self-Harness: Harnesses that improve themselves,” arXiv preprint arXiv:2606.09498, 2026.arxiv.org/abs/2606.09498
- [22]C. Yu and H. Ren, “Autonomous evolution of EDA tools: Multi-agent self-evolved ABC,” in Proceedings of the 63rd ACM/IEEE Design Automation Conference (DAC), 2026.arxiv.org/abs/2604.15082
- [23]R. Fu, Y. Liu, Q. Xu, and T.-Y. Ho, “MappingEvolve: LLM-driven code evolution for technology mapping,” in Proceedings of the 63rd ACM/IEEE Design Automation Conference (DAC), 2026.
- [24]H. Liu, L. Zhou, Y. Ren, Y. Wu, and Z. Wang, “GoalEvolve: From handcrafted algorithm priors to goal-driven evolution of physical design algorithms,” arXiv preprint arXiv:2608.16733, 2026.arxiv.org/abs/2608.16733
- [25]R. Chen, W. Xiao, R. Karri, and J. Han, “CoEvoE&R: Co-evolving placement objectives with routing feedback via large language models,” arXiv preprint arXiv:2607.17398, 2026.arxiv.org/abs/2607.17398
- [26]T. Jafri and V. A. Chhabria, “GR-Evolve: Design-adaptive global routing via LLM-driven algorithm evolution,” arXiv preprint arXiv:2604.22234, 2026.arxiv.org/abs/2604.22234
- [27]Q. Wu, T. Jafri, A. Arora, and V. A. Chhabria, “VPR-Evolve: Multi-agent-driven algorithm evolution for FPGA place and route,” arXiv preprint arXiv:2607.24998, 2026.arxiv.org/abs/2607.24998
- [28]C. Yu, C. Deng, N. Pinckney, and B. Khailany, “Agentic hardware design as repository-level code evolution,” arXiv preprint arXiv:2606.28279, 2026.arxiv.org/abs/2606.28279
- [29]K. Seyoum and A. Mittur, “Automatic harness evolution for hardware design verification: Can LLMs consolidate gains across discovered harnesses?” arXiv preprint arXiv:2609.28908, 2026.arxiv.org/abs/2609.28908
- [30]OpenAI, “GPT-5.6 Sol,” 2026, OpenAI model documentation. [Online]. Available: https://developers.openai.com/api/docs/models/gpt-5.6-sol
- [31]DeepSeek-AI, “DeepSeek-V4.1-Flash: Pushing the limits of KV cache compression,” arXiv preprint arXiv:2609.19969, 2026.
- [32]Design Compiler: RTL Synthesis Solution, Synopsys, Inc., Synopsys Design Compiler product documentation. [Online]. Available: https://www.synopsys.com/implementation-and-signoff/rtl-synthesis-test/design-compiler.html
- [33]JasperGold Sequential Equivalence Checking App, Cadence Design Systems, Inc., Cadence Jasper formal verification documentation. [Online]. Available: https://www.cadence.com/en_US/home/tools/system-design-and-verification/formal-and-static-verification/jasper-verification-platform/jaspergold-sequential-equivalence-checking-app.html
- [34]Nangate, Inc., “Nangate Open Cell Library (45 nm),” Open standard-cell library, 2008. [Online]. Available: https://eda.ncsu.edu/freepdk/freepdk45/


